Stop making your coding agent rediscover your repo.

pi-memctx gives Pi a local, durable, Markdown-native memory layer. It searches project memory before each prompt, injects only compact/relevant context, lets Pi inspect the repo normally when memory is not enough, and learns durable discoveries after turns.

No database server. No hosted memory vendor. No hidden black box. Just Markdown memory packs you can read, edit, grep, version, and sync.

Quickstart

1. Install

pi install npm:pi-memctx

Already installed? Update with:

pi update npm:pi-memctx

Or install from GitHub:

pi install git:github.com/weauratech/pi-memctx

You can also verify the local installation:

npx pi-memctx doctor

2. Start Pi with the extension

cd /path/to/your/workspace
pi -e pi-memctx

3. Generate your first memory pack

Inside Pi:

/memctx-init

pi-memctx scans the workspace and creates local Markdown workspace memory with context, decisions, runbooks, and indexes. Internally this is stored as a pack, but most users do not need to manage packs directly.

4. Ask normally

How do I deploy this service to production?

The user experience stays the same: prompt, wait, get the answer. The Memory Gateway works behind the scenes.

Why it feels like magic

A normal coding agent starts every session cold:

User: How do I deploy the gateway service to production?
Agent: I'll inspect the repo...
       [tool] list files
       [tool] read workflows
       [tool] inspect Helm charts
       [tool] search docs
       ...

pi-memctx turns that into:

User: How do I deploy the gateway service to production?
Agent: Merge to main, GitHub Actions builds Docker and pushes to ECR,
       Helm values are updated, ArgoCD syncs to Kubernetes, staging is
       automatic, and production is manual via ArgoCD approval/sync.

The difference is not a better prompt. It is memory arriving before the model starts reasoning.

Automatic learning

After a rich planning, debugging, or repository-discovery turn, pi-memctx can persist multiple linked Markdown notes instead of one shallow summary:

memctx: learned 5 memories:
   - context: [[packs/my-pack/20-context/payment-api|Payment API]] (updated)
   - observation: [[packs/my-pack/60-observations/deploy-patterns|Deploy patterns]] (created)
   - runbook: [[packs/my-pack/70-runbooks/deploy-payment-api|Deploy Payment API]] (created)
   - action: [[packs/my-pack/40-actions/2026-05-02-prepared-rollout|Prepared rollout]] (created)
   - session: [[packs/my-pack/80-sessions/rich-persistence-payment-api|Rich planning snapshot]] (created)

Learned notes are cross-linked with [[wikilinks]], so future searches can recover the whole discovery: context, observations, runbooks, actions, decisions, and rich session snapshots.

Local and inspectable by design

• Memories are Markdown files on your machine.
• No hosted memory service or external vector database is required.
• You can inspect, edit, delete, commit, or sync packs yourself.
• Secret-looking values are blocked/redacted before persistence.
• When memory is insufficient or stale, Pi falls back to normal repo inspection.

Benchmark

Real-world stress benchmark

A local stress benchmark was run against five anonymized real workspace-memory packs (Pack 1 … Pack 5), with 5 repeats per task and two profiles: plain Pi with no extensions (baseline) and Pi with pi-memctx (gateway). The benchmark is read-only: no commits, deploys, applies, publishes, or cloud mutations.

| Profile | Runs | Avg latency | Provider tokens/task | Visible tokens/task | Tool calls/task | Quality | Timeouts | |---|---:|---:|---:|---:|---:|---:|---:| | baseline | 25 | 196.6s | 34,385 | 2,503 | 27.2 | 206/255 | 1 | | gateway | 25 | 65.0s | 6,614 | 1,776 | 1.36 | 231/255 | 0 |

Compared with baseline:

| Metric | Gateway vs baseline | |---|---:| | Latency | 66.9% faster | | Provider tokens | 80.8% fewer | | Tool calls | 95.0% fewer | | Quality | +25 facts | | Timeouts | 1 → 0 |

Per-pack quality in the same anonymized run:

| Case | Baseline | Gateway | |---|---:|---:| | Pack 1 — deployment/GitOps runbook recall | 56/60 | 60/60 | | Pack 2 — Lambda/release workflow recall | 48/50 | 45/50 | | Pack 3 — package publishing configuration | 44/45 | 38/45 | | Pack 4 — long frontend-workflow continuation | 30/50 | 38/50 | | Pack 5 — observability troubleshooting runbook | 28/50 | 50/50 |

The biggest change is not just quality; it is avoiding expensive rediscovery. The gateway keeps average tool use near one call per prompt while still allowing bounded fallback when memory is partial.

Reproducible synthetic benchmark

The repository also includes a synthetic benchmark fixture you can run locally:

bash benchmark/setup.sh /tmp/pi-memctx-benchmark
BENCH_REPEATS=2 \
BENCH_PROFILES="baseline gateway" \
bash benchmark/run.sh /tmp/pi-memctx-benchmark

Recent synthetic release-check run (BENCH_REPEATS=5, 25 tasks/profile, public fixture only):

| Profile | Runs | Avg latency | Provider tokens/task | Visible tokens/task | Tool calls/task | Failed tools/task | Quality | |---|---:|---:|---:|---:|---:|---:|---:| | baseline | 25 | 22.0s | 2,450 | 571 | 6.2 | 0.4 | 63/110 | | gateway | 25 | 10.5s | 2,615 | 329 | 0.3 | 0.0 | 76/110 |

Gateway result on the public fixture:

• 52.2% faster average latency.
• 95.2% fewer tool calls.
• 42.4% fewer visible tokens in assistant-facing prompt/output text.
• +13 scored facts across the same prompts.
• Provider token totals may move differently from visible tokens because model/provider cache accounting is included.

Benchmarks are intentionally local and reproducible. Results depend on the model/provider, machine, cache behavior, and contents of your memory packs.

How it works

pi-memctx is a Memory Gateway between the user and the main LLM.

User prompt
   │
   ▼
Memory Gateway
   ├─ detects the active pack, including workspace maps and prompt mentions
   ├─ supports Markdown packs that are real directories or symlinked directories
   ├─ retrieves relevant Markdown memories with qmd or grep fallback
   ├─ scopes qmd collections by pack path to avoid cross-vault index reuse
   ├─ ranks candidates with cheap semantic coverage
   ├─ builds a compact local memory summary
   └─ injects only useful context
   │
   ▼
Pi agent answers normally
   ├─ uses memory when sufficient
   ├─ inspects the repo when memory is partial or stale
   └─ after the turn, pi-memctx learns durable context back to Markdown
      ├─ context
      ├─ observations
      ├─ runbooks
      ├─ decisions
      ├─ actions
      └─ rich session snapshots

The gateway does not replace the main LLM. It does the boring part first: finding the right project memory, compressing it, and preventing redundant tool exploration when the answer is already known.

When multiple packs are available, explicit pack names in the prompt take priority over weaker aliases from other packs. This keeps multi-workspace setups predictable without hardcoded company, domain, language, or technology rules.

Profile

pi-memctx now has one recommended runtime profile: gateway.

| Profile | Best for | Behavior | |---|---|---| | gateway | Fast daily use | Conservative local judge, compact context, qmd/grep retrieval, zero redundant memory searches when memory is sufficient. |

The profile is applied by default. Use /memctx-status --advanced if you need to inspect the active runtime settings. Old profile names such as gateway-lite, gateway-full, qmd-economy, low, balanced, auto, and full are compatibility-mapped to gateway in configuration files.

Workspace memory

Each workspace gets local Markdown memory. Internally, workspace memory is stored as a pack directory with frontmatter. You can edit it with any editor, commit it, review it, or open it in Obsidian.

packs/my-project/
  00-system/
    pi-agent/
      memory-manifest.md
      resource-map.md
    indexes/
      context-index.md
      decision-index.md
      runbook-index.md
  20-context/
    api.md
    web.md
    infra.md
  40-actions/
    2026-05-02-prepared-rollout.md
  50-decisions/
    001-hexagonal-architecture.md
    002-use-pgx.md
  60-observations/
    deploy-patterns.md
  70-runbooks/
    deploy.md
    terraform.md
  80-sessions/
    rich-persistence-payment-api.md

Recommended note types:

| Type | Directory | Use it for | |---|---|---| | context | 20-context/ | Stack, services, repositories, conventions, environments. | | decision | 50-decisions/ | Architecture and technical decisions with rationale. | | observation | 60-observations/ | Durable facts, requirements, caveats, structural discoveries. | | runbook | 70-runbooks/ | Repeatable operational procedures. | | action | 40-actions/ | Completed work, migrations, deploys, incident notes. | | session | 80-sessions/ | Sanitized rich planning/discovery snapshots for future retrieval. |

Commands

Most users only need the daily commands:

| Command | Purpose | |---|---| | /memctx | Show compact help and current memory status. | | /memctx-init | Create/update memory for this workspace. If a model is selected, deep LLM enrichment starts in the background by default; use --no-deep to skip it. | | /memctx-status | Show workspace memory status; add --advanced for internal paths/config. | | /memctx-review | Review queued memory candidates; approve, reject, clear, or inspect the queue path. | | /memctx-refresh | Refresh workspace memory inventory/enrichment in the background. | | /memctx-doctor | Diagnose qmd, workspace memory, and configuration. |

Older pack/config commands were removed from the public command surface to keep the extension simple. Advanced behavior is still configurable through environment variables and the local config file.

Reviewing queued memory candidates

When automatic learning is unsure whether a candidate should be persisted, pi-memctx queues it for review instead of silently saving it. Review the active workspace queue with:

/memctx-review

Useful forms:

/memctx-review --list
/memctx-review approve 1
/memctx-review reject 1
/memctx-review clear
/memctx-review path

By default, review commands operate only on the active workspace memory pack. Advanced users can inspect the raw queue at ~/.cache/pi-memctx/save-queue.json.

Tools

memctx_search

Search the active memory pack:

Use memctx_search to find the deploy runbook.

Parameters:

| Parameter | Values | Default | |---|---|---| | query | string | required | | mode | keyword, semantic, deep | keyword | | limit | number | 5 |

When the Memory Gateway already injected sufficient memory, the agent is instructed not to call memctx_search again. That keeps answers fast and prevents duplicate context retrieval.

Gateway diagnostics

For memory-quality investigations, enable opt-in debug snapshots:

MEMCTX_GATEWAY_DEBUG=1 pi -e pi-memctx

Snapshots are written under:

$TMPDIR/pi-memctx-gateway-debug/

They include sanitized/truncated prompt terms, active pack, candidate paths, selected facts, coverage, fallback budget, and top scored candidate lines. Debug is off by default, best-effort only, and intended for local diagnosis; do not commit debug snapshots because they may contain project-specific paths or note titles.

memctx_save

Save durable knowledge into the active pack:

Remember that production deploys require ArgoCD manual approval.

The tool supports:

• observation
• decision
• action
• runbook
• context
• session

Secret-looking content is blocked.

Status overlay

pi-memctx keeps a small status overlay in Pi:

🧠 my-pack · memory ready · 3 memory hits · qmd · learn auto

This tells you which pack is active, whether memory was useful, which search backend is being used, and whether automatic learning is enabled.

Best use cases

pi-memctx is especially useful when:

• you work on large repos or monorepos;
• Pi keeps rediscovering architecture and deploy flows;
• your team has many services with repeated conventions;
• you want local memory without a hosted vector database;
• you want agent memory you can read, edit, grep, and version;
• you want planning/discovery sessions to become durable project notes.

Why not just RAG?

Most RAG setups retrieve documents at query time. pi-memctx is different:

• it is local-first and file-based;
• it stores durable memories as Markdown;
• it learns after turns, not only before prompts;
• it links related memories together;
• it can fall back to repo inspection when memory is stale;
• it does not require a hosted vector database.

qmd integration

pi-memctx uses @tobilu/qmd when available for fast memory retrieval. To keep pi install npm:pi-memctx clean and warning-free, qmd is not installed automatically. If qmd is not available, pi-memctx falls back to grep-based search.

Resolution order:

1. QMD_PATH or MEMCTX_QMD_BIN
2. an already-installed local .bin or bundled/vendor path
3. qmd on PATH
4. grep fallback

Optional qmd install example:

npm install -g @tobilu/qmd

Check your setup:

npx pi-memctx doctor

Safety model

pi-memctx is local-first and intentionally boring about sensitive data.

It will not save:

• API keys
• tokens
• passwords
• private keys
• customer data
• payment card data
• sensitive payloads

Memory is Markdown on disk. You can inspect every byte.

Configuration

Most users should start with the defaults. Advanced users can configure behavior with environment variables or the local config file at ~/.config/pi-memctx/config.json.

Common environment variables:

| Variable | Purpose | |---|---| | MEMCTX_CONTEXT_TOKEN_BUDGET | Approximate injected context budget. | | MEMCTX_CONTEXT_MAX_ITEMS | Maximum memory items to include. | | MEMCTX_RETRIEVAL | Retrieval policy: fast, balanced, deep, strict, auto. | | MEMCTX_GATEWAY_JUDGE | Gateway judge mode: off, conservative, auto, main-llm. | | MEMCTX_AUTOSAVE | Autosave mode: off, suggest, confirm, auto. | | QMD_PATH / MEMCTX_QMD_BIN | Explicit qmd binary path. |

Development

Requirements:

• Node.js 20+
• Bun for tests
• Pi coding agent

Install dependencies:

npm install

Run checks:

npm run typecheck
npm test
npm run test:e2e

Run all CI checks:

npm run ci

Run benchmark:

bash benchmark/setup.sh /tmp/pi-memctx-benchmark
QMD_PATH=/tmp/pi-memctx-qmd/node_modules/.bin/qmd \
BENCH_REPEATS=1 \
BENCH_PROFILES="baseline gateway" \
BENCH_TIMEOUT=120 \
bash benchmark/run.sh /tmp/pi-memctx-benchmark

Contributing

Issues and pull requests are welcome.

Good contributions include:

• better language-agnostic retrieval heuristics
• safer memory extraction
• clearer benchmark fixtures
• docs and examples
• profile tuning with reproducible numbers

Please do not commit private company memory packs, customer data, secrets, or benchmark fixtures derived from private repositories.

License

MIT